[iccrg] Prague improvements (was RE: [tsvwg] [CCWG] New Internet Draft: Congestion Signaling (CSIG))

[Ingemar Johansson S <[e-mail]>]

Hi Bob + others

Interesting to hear about PragueJ. The paper was somehow below radar coverage.  
We see a bit of an issue with the normal additive increase when we experiment with mmWave access (only simulations) where blockage to a smaller or greater extent can cause the link throughput to drop dramatically. Normal additive increase is quite slow to converge to e.g 2Gbps throughput, even though the RTT is low. 

To overcome the problem I implemented in Prague a multiplicative increase of CWND where the CWND can increase by up to for instance 5% of the CWND per RTT. The degree of multiplicative increase is scaled based on the number of rounds since the last congestion event. This makes the rate converge considerably faster but there is a certain amount of overshoot.
The delay detection that you suggest can solve this issue. One possible challenge I can see is that wireless access can have delay jitter not only due to retransmissions on the MAC layer but also on the RLC later, although the latter is typically less common. How would the delay option handle such outliers.

That said.. It would be very good to come up with improved scalable congestion controls that exploit the potential with L4S better.

/Ingemar

 

> -----Original Message-----
> From: Bob Briscoe <[e-mail]>
> Sent: Monday, 18 September 2023 23:38
> To: Toerless Eckert <[e-mail]>
> Cc: K. K. Ramakrishnan <[e-mail]>; Christian Huitema
> <[e-mail]>; iccrg IRTF list <[e-mail]>
> Subject: Re: [iccrg] [tsvwg] [CCWG] New Internet Draft: Congestion Signaling
> (CSIG)
> 
> Toerless,
> 
> 
> On 18/09/2023 19:43, Toerless Eckert wrote:
> > Nice explanation, Bob, thanks.
> >
> > I guess there is also a discovery of a non-ECN capable congestion
> > point along the path and fall-back to a less aggressive (non scalable)
> signaling ?
> 
> [BB] Currently, Prague is only more responsive (less aggressive) to loss ,
> but just for one response - there is no 'mode change' longer term.
> 
> Nonetheless, Joakim's experimental modifications to Prague (called PragueJ
> from here on) are more responsive to delay too. But, like the acceleration
> mode, the response to delay is a one-off overload response, not using delay
> as a continual congestion avoidance signal, as Vegas or BBR do. If PragueJ
> sees both ECN and delay, it does a once-off 'overload' response aiming to
> remove all the delay (and no more than
> all) in one RTT.
> 
> > What is todays best timing of a) discovering that, and b)
> > undiscovering that
> > (aka: returning to ECN control once the non-ECN hop is not a
> > congestion point anymore).
> 
> For the overload response, the discovery threshold is not primarily time,
> just an increase in delay, altho it does use an EWMA of delay with a
> characteristic time of less than a round trip.
> 
> The discovery / undiscovery time before entering or leaving acceleration
> mode are particularly interesting.
> 
> Discovery: PragueJ keeps EWMAs of the round trips between ECN marks and of
> the absolute deviation. Then it starts acceleration mode after multiples of
> these EWMAs. So the discovery time is not hard-coded, but depends on a
> departure from the recent norm of the environment. {Note 1}
> 
> Undiscovery: The EWMAs continue to be maintained during the acceleration,
> then the variables being averaged (but not the averages
> themselves) are reset once an ECN signal is received. This ensures that the
> algorithm can only enter acceleration mode again once there has been a long
> enough period of stability, taking account of how much it moved from its
> previous period of stability (see §3 for details).
> 
> Finally, the acceleration itself is nice. During acceleration mode, PragueJ
> does additive increase by 'a' segments, where 'a' is calculated so that the
> overshoot after one RTT of feedback delay cannot cause greater than a
> certain amount of queue delay, called Q in the paper (e.g. 1ms).
> 
> Each flow scales its additive increase assuming the bottleneck link's
> capacity is filled solely by its own window. So, with multiple flows, the
> delay due to all the overshoots added together still cannot exceed Q.
> 
> 
> Note 1: (Joakim and I developed a principle where we try to measure all
> first-order parameters and only use 'magic numbers'  as second-order
> parameters to control the dynamics of the measurements that collect the
> first-order parameters.)
> 
> Cheers
> 
> 
> Bob
> 
> >
> > Cheers
> >      Toerles
> >
> > On Mon, Sep 18, 2023 at 07:04:28PM +0100, Bob Briscoe wrote:
> >> KK, Christian,
> >>
> >> I want to focus for a moment on the problem of how a flow in progress
> >> knows that available capacity has increased (e.g. because other
> >> flow(s) have departed or bottleneck link capacity has increased).
> >> Current congestion control algorithms (CCAs) take so long to detect
> >> an absence of signals (loss, delay or ECN), that they use the same
> >> probing algorithm that they use in their congestion avoidance phase to
> fill any available capacity.
> >>
> >> With Reno, Cubic, etc. the time between congestion episodes in steady
> >> state for a single flow is now hundreds of RTT. So, it takes even
> >> more than that before you can be sure there is a real absence of
> congestion signals.
> >>
> >> Scalable CCAs do a lot better. The target average time between
> >> signals in stable conditions is half an RTT. You can sustain such a
> >> high signalling rate with ECN, but you couldn't do it with loss,
> >> which is why L4S is tightly associated with ECN. So, after just a few
> >> RTT of no signals you can start probing for more capacity more
> >> aggressively than you would want to in congestion avoidance phase.
> >>
> >> This is the critical point - the CCAs of all flows that notice the
> >> absence of signals briefly enter a separate acceleration mode where
> >> convergence is a non-goal. Then, when they find a closed loop signal
> >> again, they return to their congestion avoidance mode to converge
> >> with each other from wherever they got to when they 'broke loose'.
> >>
> >> Before Joakim Misund decided to move on from his doctoral research
> >> into getting up to speed fast, he did some nice experiments on this
> >> acceleration mode.
> >> For instance, look at Fig 5 in this draft tech report he wrote:
> >> https://bobbriscoe.net/projects/latency/04-delay_bound_faster_additiv
> >> e_increase_when_unterutilize_tr.pdf
> >>
> >> In Fig 5 capacity alternates between 20 Mb/s and 200 Mb/s in a square
> wave.
> >> It compares his modified TCP Prague (the pink plot) with BBRv2 &
> >> unmodified Prague.  The legend calls the pink plot "Prague w/both"
> >> because it has an acceleration mode and an overload mode to rapidly
> >> decrease when there is both ECN and heavy delay.
> >>
> >>
> >> Bob
> >>
> >>
> >> On 16/09/2023 12:13, K. K. Ramakrishnan wrote:
> >>> I agree. This was the original idea  for a "faster" ramp up
> >>> initially, Christian. It is (if available on the end-end path) a
> >>> relatively benign solution that doesn't stomp out existing flows.
> >>> If ECN is not available on the path, it is easy enough to not ramp up.
> >>>
> >>> On Fri, Sep 15, 2023 at 9:19 PM Christian Huitema
> >>> <[e-mail]>
> >>> wrote:
> >>>
> >>>      RFC 9049 is an interesting read. Reading it, I was nodding my
> head...
> >>>      until it went to the ECN section. The big change we are seeing is
> >>>      that
> >>>      deploying ECN now is mostly safe -- QUIC, for example, can test for
> >>>      end-to-end availability of ECN, and measurements show that it is
> >>>      available on many paths. It may not be actively implemented by all
> >>>      elements in the path, but at least nothing crashes, and we can see
> >>>      deployments. It only took two decades to get there.
> >>>
> >>>      Which means that maybe we should start fully exploiting ECN for
> >>>      what it
> >>>      is worth. Probably not a complete solution to the "quick start"
> >>>      problem,
> >>>      but at least a partial solution, one that would complement packet
> >>>      loss
> >>>      detection and delay increase measurements. That seems more
> promising
> >>>      than trying to deploy another solution across all Internet routers.
> >>>
> >>>      And maybe we can make creative use of ECN, even in cases of long
> >>>      delay
> >>>      links. Combine it with chirping, maybe as in "send a chirp at a
> quick
> >>>      rate, and check whether that chirp elicited ECN feedback?"
> >>>
> >>>      -- Christian Huitema
> >>>
> >>>      On 9/15/2023 1:58 AM, Gorry Fairhurst wrote:
> >>>      > I wonder if RFC 9049 - Path Aware Networking: Obstacles to
> >>>      Deployment (A
> >>>      > Bestiary of Roads Not Taken), section 6.3, might have some
> thoughts
> >>>      > relevant to why some approaches in DC research diverge from the
> >>>      > end-to-end approaches described in published RFCs.
> >>>      >
> >>>      > Best wishes,
> >>>      > Gorry
> >>>      >
> >>>      >
> >>>      > On 15/09/2023 08:37, Hesham ElBakoury wrote:
> >>>      >> Speaking of Internet congestion control, have you looked at this
> >>>      >> paper: Future Internet Congestion Control: The Diminishing
> >>>      Feedback
> >>>      >> Problem [https://arxiv.org/pdf/2206.06642]? The abstract of
> >>>      this paper
> >>>      >> says: /<we argue that congestion control mechanisms should move
> >>>      away
> >>>      >> from their strict “end-to-end”/
> >>>      >> /adherence. This change would benefit from avoiding a “one size
> >>>      fits
> >>>      >> all circumstances” approach, and moving towards a more/
> >>>      >> /selective set of mechanisms that will result in a better
> >>>      performing
> >>>      >> Internet./>
> >>>      >>
> >>>      >> Hesham
> >>>      >>
> >>>      >> On Thu, Sep 14, 2023, 3:23 AM Sebastian Moeller
> >>>      <[e-mail]> wrote:
> >>>      >>
> >>>      >>     Hi Ruediger,
> >>>      >>
> >>>      >>
> >>>      >>     > On Sep 14, 2023, at 09:25, <[e-mail]>
> >>>      >>     <[e-mail]> wrote:
> >>>      >>     >
> >>>      >>     > Sebastian, Ingemar,
> >>>      >>     >
> >>>      >>     > the draft says "limited domain". The method suggested
> >>>      resembles
> >>>      >>     IPPM's IOAM (but the packet size remains fixed, certainly an
> >>>      >>     advantage). Information like a port ID only makes sense, if
> the
> >>>      >>     evaluation unit is aware of the domain-topology.
> >>>      >>
> >>>      >>             [SM] Not sure I understand this, if on an internet-
> wide
> >>>      >>     path all nodes would simply include the minimum of their
> egress
> >>>      >>     bitrate (that is only include this if it is lower than the
> >>>      >>     existing value), then an end-2end protocol like TCP could
> >>>      even as
> >>>      >>     part of the initial 3 way handshake figure out a hard limit
> for
> >>>      >>     maximum rasinsable cwin (by simply calculating the BDP).
> >>>      SUre that
> >>>      >>     would not stop slow-start fuully from overshooting, but it
> >>>      should
> >>>      >>     limit the maximum overshoot considerably....
> >>>      >>
> >>>      >>
> >>>      >>     > That will be restricted to a single domain, I think.
> >>>      >>
> >>>      >>             [SM] As long as the forward and collection path
> >>>      requires
> >>>      >>     L2 elements, sure this will not work over the internet.
> >>>      >>
> >>>      >>
> >>>      >>     > I'm also not sure, what is to be optimized on an
> operational
> >>>      >>     carrier backbone.
> >>>      >>
> >>>      >>             [SM] I trust you on this, as that is out of my
> league.
> >>>      >>
> >>>      >>
> >>>      >>     > Under expectable conditions, these are operated without
> >>>      >>     congestion. Bottlenecks are likely at the access and, in a
> >>>      rather
> >>>      >>     limited number of cases, at peering. There you'd need to
> >>>      know the
> >>>      >>     RTT so that the available bandwidth calculation makes sense.
> >>>      >>
> >>>      >>             [SM] But the protocol that is supposed to consume
> the
> >>>      >>     congestion information will know the RTT, so can take this
> into
> >>>      >>     account, the network really only needs to give information
> >>>      about
> >>>      >>     its current state. That said, I can see network nodes that
> >>>      might
> >>>      >>     want to know a flow's RTT, but I do not think there is a way
> of
> >>>      >>     supplying that this is not both easy and efficient and not
> >>>      easily
> >>>      >>     gamed. Passing on congestion informatin however is far less
> >>>      >>     abusable (after all the node could just as well have
> >>>      dropped the
> >>>      >>     packet instead of bothering to fudge the congestion info).
> >>>      >>
> >>>      >>
> >>>      >>     > Or you need to calculate several values for several RTTs
> and
> >>>      >>     provide this information (RTTs and corresponding available
> >>>      >>     bandwidths). Requires a fair amount of bytes, so then trust
> >>>      may be
> >>>      >>     an issue (I'm not sure whether a carrier would be willing
> >>>      to add
> >>>      >>     an extended Ethernet header to every encrypted IP packet
> >>>      >>     indicating interest in getting CSIG information on
> >>>      bottleneck links).
> >>>      >>
> >>>      >>             [SM] Yes encryption/privacy are reasons to stick to
> >>>      L2/L3.
> >>>      >>     I would guess in a limited domain deploying CSIG as a
> >>>      pseudo VLAN
> >>>      >>     header seems like a nice work-around, but this information
> >>>      really
> >>>      >>     should be inside IP header to go end to end. IMHO it also
> >>>      should
> >>>      >>     not be in an optional header, but should have been part of
> the
> >>>      >>     mandatory header, but that ship has sailed long ago... It
> turns
> >>>      >>     out that the mechanism intended here IPv6 extension headers
> was
> >>>      >>     mis-deployed... (It should have been initiated with a really
> >>>      >>     desirable use-case so it would actrually see usage
> immediately,
> >>>      >>     protocol design is a bit like evolution, in both cases "use
> >>>      it or
> >>>      >>     loose it" applies to some degree.)
> >>>      >>
> >>>      >>     Regards
> >>>      >>             Sebastian
> >>>      >>
> >>>      >>
> >>>      >>     >
> >>>      >>     > Regards, Ruediger
> >>>      >>     >
> >>>      >>     > -----Ursprüngliche Nachricht-----
> >>>      >>     > Von: Sebastian Moeller <[e-mail]>
> >>>      >>     > Gesendet: Donnerstag, 14. September 2023 08:49
> >>>      >>     > An: Ingemar Johansson S <[e-mail]>
> >>>      >>     > Cc: Geib, Rüdiger <[e-mail]>;
> [e-mail];
> >>>      >> [e-mail]; [e-mail]; [e-mail]
> >>>      >>     > Betreff: Re: [tsvwg] [iccrg] New Internet Draft:
> Congestion
> >>>      >>     Signaling (CSIG)
> >>>      >>     >
> >>>      >>     > Hi Ingemar
> >>>      >>     >
> >>>      >>     >> On Sep 14, 2023, at 08:42, Ingemar Johansson S
> >>>      >>     <[e-mail]> wrote:
> >>>      >>     >>
> >>>      >>     >> Hi Ruediger
> >>>      >>     >>
> >>>      >>     >> Even though CSIG is on ethernet, it appears to be e2e as
> the
> >>>      >>     feedback is on L4. So I guess somehow the CSIG info needs
> >>>      to jump
> >>>      >>     from domain to domain somewhow, and that sounds to me like
> L3,
> >>>      >>     albeit perhaps brief jumps ?
> >>>      >>     >
> >>>      >>     >       [SM] rfc3168 and L4S ECN signaling faces the same
> >>>      hurdle,
> >>>      >>     forward path uses a different layer than feed-back path.
> >>>      And I do
> >>>      >>     not think that it is conceptually all that cleaner to have
> >>>      fewer
> >>>      >>     layers between forward and reverse signaling... "clean"
> >>>      would be
> >>>      >>     to put both forward and reverse information into the same
> >>>      layer,
> >>>      >>     but that is not on the table as far as I can see.
> >>>      >>     >       The other difference however, richness of signal, is
> a
> >>>      >>     clear difference between 1 bit ECN and multibit CSIG (and
> >>>      >>     alternatives). (Side-note, sure the ECN bitfield is two
> >>>      bits but
> >>>      >>     it only carries 2 congestion states when ECN is in use,
> which
> >>>      >>     makes it IMHO a 1-bit information channel, if we had
> >>>      followed the
> >>>      >>     SCE proposal with its 3 congestion states we would be at
> >>>      1.5 bits ;))
> >>>      >>     >
> >>>      >>     > Regards
> >>>      >>     >       Sebastian
> >>>      >>     >
> >>>      >>     >
> >>>      >>     >
> >>>      >>     >>
> >>>      >>     >> /Ingemar
> >>>      >>     >>
> >>>      >>     >>> -----Original Message-----
> >>>      >>     >>> From: [e-mail]
> <[e-mail]>
> >>>      >>     >>> Sent: Wednesday, 13 September 2023 14:20
> >>>      >>     >>> To: [e-mail]; Ingemar Johansson S
> >>>      >>     >>> <[e-mail]>
> >>>      >>     >>> Cc: [e-mail]; [e-mail];
> >>>      [e-mail];
> >>>      >>     >>> [e-mail]; [e-mail];
> [e-mail];
> >>>      >>     >>> [e-mail]; [e-mail];
> [e-mail];
> >>>      >>     >>> [e-mail]; [e-mail];
> [e-mail]
> >>>      >>     >>> Subject: AW: [tsvwg] [iccrg] New Internet Draft:
> Congestion
> >>>      >>     Signaling
> >>>      >>     >>> (CSIG)
> >>>      >>     >>>
> >>>      >>     >>> Hi Ingemar, hi Sebastian,
> >>>      >>     >>>
> >>>      >>     >>> Skimming over the draft only, isn't CSIG about
> >>>      Ethernet-domains,
> >>>      >>     >>> while L4S is E2E?
> >>>      >>     >>>
> >>>      >>     >>> Regards,
> >>>      >>     >>>
> >>>      >>     >>> Ruediger
> >>>      >>     >>>
> >>>      >>     >>> -----Ursprüngliche Nachricht-----
> >>>      >>     >>> Von: tsvwg <[e-mail]> Im Auftrag von
> >>>      Sebastian
> >>>      >>     Moeller
> >>>      >>     >>> Gesendet: Mittwoch, 13. September 2023 13:42
> >>>      >>     >>> An: Ingemar Johansson S
> >>>      >>     >>> <ingemar.s.johansson=[e-mail]>
> >>>      >>     >>> Cc: Vidhi Goel <vidhi_goel=[e-mail]>;
> >>>      >>     >>> Shihang(Vincent)
> >>>      <shihang9=[e-mail]>; tsvwg
> >>>      >>     >>> <[e-mail]>; Jai Kumar <[e-mail]>;
> IETF
> >>>      >>     IPPM WG
> >>>      >>     >>> <[e-mail]>; Tom Herbert
> >>>      >>     <tom=[e-mail]>;
> >>>      >>     >>> [e-mail]; Abhiram Ravi
> >>>      >>     <abhiramr=[e-mail]>;
> >>>      >>     >>> Nandita Dukkipati <[e-mail]>; [e-mail];
> >>>      >>     Huangyihong
> >>>      >>     >>> (Rachel) <rachel.huang=[e-mail]>;
> >>>      Naoshad
> >>>      >>     Mehta
> >>>      >>     >>> <[e-mail]>
> >>>      >>     >>> Betreff: Re: [tsvwg] [iccrg] New Internet Draft:
> Congestion
> >>>      >>     Signaling
> >>>      >>     >>> (CSIG)
> >>>      >>     >>>
> >>>      >>     >>> Hi Ingemar,
> >>>      >>     >>>
> >>>      >>     >>>
> >>>      >>     >>>> On Sep 13, 2023, at 12:30, Ingemar Johansson S
> >>>      >>     >>> <ingemar.s.johansson=[e-mail]>
> wrote:
> >>>      >>     >>>>
> >>>      >>     >>>> Hi
> >>>      >>     >>>>
> >>>      >>     >>>> I agree with Vihdi
> >>>      >>     >>>>
> >>>      >>     >>>> L4S is recently standardised
> >>>      >>     >>>
> >>>      >>     >>>     [SM] In experimental track, the goal is currently
> >>>      to test
> >>>      >>     whether it
> >>>      >>     >>> can/should be deployed at scale....
> >>>      >>     >>>
> >>>      >>     >>>> and it is definitely gaining traction also in 3GPP. We
> >>>      have
> >>>      >>     an echo
> >>>      >>     >>> system that is looking forward to having L4S widely
> >>>      deployed.
> >>>      >>     >>>> Still, the congestion control aspects are not fully
> >>>      explored
> >>>      >>     yet.
> >>>      >>     >>>> One
> >>>      >>     >>> interesting topic is if L4S allows to more safely
> >>>      deviate from
> >>>      >>     >>> additive increase to make congestion control algorithms
> >>>      more
> >>>      >>     quickly
> >>>      >>     >>> converge to higher link capacity. There are a number of
> >>>      study
> >>>      >>     topics
> >>>      >>     >>> around L4S congestion control that are listed in e.g
> >>>      the TCP
> >>>      >>     Prague draft.
> >>>      >>     >>>>
> >>>      >>     >>>> I cannot dictate what others should do with their time
> and
> >>>      >>     money but
> >>>      >>     >>> personally I'd prefer that the IETF explores L4S and its
> >>>      >>     >>> possibilities and downsides before jumping on the next
> >>>      idea.
> >>>      >>     >>>
> >>>      >>     >>>     [SM] L4S can be described as taking the ideas
> >>>      behind DCTCP
> >>>      >>     and
> >>>      >>     >>> making them fit for use over the internet*. Yet the
> >>>      signaling
> >>>      >>     >>> discussed here is to be used in e.g. data center
> contexts
> >>>      >>     where DCTCP
> >>>      >>     >>> is already used and found lacking compared to newer
> methods
> >>>      >>     operating
> >>>      >>     >>> on richer congestion information (HPCC, Swift,
> >>>      Poseidon, ...).
> >>>      >>     >>>     Given that L4S essentially uses a multi-packet
> >>>      signal(**)
> >>>      >>     to report
> >>>      >>     >>> the "queue filling state" that is then stochastically
> >>>      distributed
> >>>      >>     >>> over all concurrent flows, it seems obvious to me that
> >>>      >>     reconstructing
> >>>      >>     >>> a reliable estimate of on-path queueing will take some
> >>>      time and
> >>>      >>     >>> averaging for each individual flow, I would guess that
> >>>      in some
> >>>      >>     >>> environments this delay simply is too costly.
> >>>      >>     >>>     So L4S and CSIG seem complementary and in no way
> >>>      mutually
> >>>      >>     exclusive.
> >>>      >>     >>>
> >>>      >>     >>>
> >>>      >>     >>> Regards
> >>>      >>     >>>     Sebastian
> >>>      >>     >>>
> >>>      >>     >>>
> >>>      >>     >>>
> >>>      >>     >>> *) I will not further discuss whether that is achieved
> >>>      or not
> >>>      >>     as it
> >>>      >>     >>> seems irrelevant here.
> >>>      >>     >>> **) In essence transmission of congestion state via a 1-
> bit
> >>>      >>     serial
> >>>      >>     >>> channel, clocked at the (variable***) packet rate at the
> >>>      >>     bottleneck.
> >>>      >>     >>> ***) as packets are not of uniform size
> >>>      >>     >>>
> >>>      >>     >>>>
> >>>      >>     >>>> CSIG sounds to me like something that belongs more in
> >>>      ICCRG or ?
> >>>      >>     >>>>
> >>>      >>     >>>> /Ingemar
> >>>      >>     >>>>
> >>>      >>     >>>> From: tsvwg <[e-mail]> On Behalf Of
> >>>      Vidhi Goel
> >>>      >>     >>>> Sent: Wednesday, 13 September 2023 00:59
> >>>      >>     >>>> To: Shihang(Vincent)
> >>>      <shihang9=[e-mail]>
> >>>      >>     >>>> Cc: Huangyihong (Rachel)
> >>>      >>     <rachel.huang=[e-mail]>;
> >>>      >>     >>> Tom Herbert <tom=[e-mail]>;
> >>>      Abhiram Ravi
> >>>      >>     >>> <abhiramr=[e-mail]>; IETF IPPM WG
> >>>      >>     <[e-mail]>;
> >>>      >>     >>> tsvwg <[e-mail]>; [e-mail]; [e-mail];
> >>>      Nandita
> >>>      >>     >>> Dukkipati <[e-mail]>; Naoshad Mehta
> >>>      >>     <[e-mail]>;
> >>>      >>     >>> Jai Kumar <[e-mail]>
> >>>      >>     >>>> Subject: Re: [tsvwg] [iccrg] New Internet Draft:
> >>>      Congestion
> >>>      >>     >>>> Signaling
> >>>      >>     >>> (CSIG)
> >>>      >>     >>>>
> >>>      >>     >>>> Not sure why we are coming up with so many new
> techniques
> >>>      >>     when ECN
> >>>      >>     >>> just works fine.
> >>>      >>     >>>> ECN is a 2 bit field (not 1 bit) and seems to be
> >>>      sufficient to
> >>>      >>     >>> indicate extent of congestion by marking it per packet.
> >>>      Adding
> >>>      >>     more
> >>>      >>     >>> complexity to any layer whether it is L2 or L3 doesn’t
> work
> >>>      >>     well in
> >>>      >>     >>> deployments. Our goal should be to simplify things and
> only
> >>>      >>     add new
> >>>      >>     >>> headers if absolutely necessary.
> >>>      >>     >>>>
> >>>      >>     >>>> Vidhi
> >>>      >>     >>>>
> >>>      >>     >>>>
> >>>      >>     >>>> On Sep 12, 2023, at 3:12 AM, Shihang(Vincent)
> >>>      >>     >>> <shihang9=[e-mail]> wrote:
> >>>      >>     >>>>
> >>>      >>     >>>> Hi,
> >>>      >>     >>>> I agree L2 may not be the best choice to carry the
> >>>      congestion
> >>>      >>     >>> signaling end-to-end and more bits are needed. We have
> >>>      >>     submitted a
> >>>      >>     >>> draft to carry the multi-bits congestion signaling in
> >>>      L3. We
> >>>      >>     call it
> >>>      >>     >>> Advanced ECN. See
> >>>      >> https://datatracker.ietf.org/doc/draft-shi-ccwg-advanced-ecn/.
> >>>      >>     >>>>
> >>>      >>     >>>> Thanks,
> >>>      >>     >>>> Hang
> >>>      >>     >>>>
> >>>      >>     >>>> From: CCWG <[e-mail]> On Behalf Of
> >>>      Huangyihong
> >>>      >>     (Rachel)
> >>>      >>     >>>> Sent: Tuesday, September 12, 2023 5:41 PM
> >>>      >>     >>>> To: Tom Herbert <tom=[e-mail]>;
> >>>      >>     Abhiram Ravi
> >>>      >>     >>> <abhiramr=[e-mail]>
> >>>      >>     >>>> Cc: IETF IPPM WG <[e-mail]>; tsvwg
> <[e-mail]>;
> >>>      >>     >>> [e-mail]; [e-mail]; Nandita Dukkipati
> >>>      >>     >>> <[e-mail]>; Naoshad Mehta
> >>>      <[e-mail]>; Jai
> >>>      >>     Kumar
> >>>      >>     >>> <[e-mail]>
> >>>      >>     >>>> Subject: Re: [CCWG] [iccrg] [tsvwg] New Internet Draft:
> >>>      >>     Congestion
> >>>      >>     >>> Signaling (CSIG)
> >>>      >>     >>>>
> >>>      >>     >>>> Hi,
> >>>      >>     >>>>
> >>>      >>     >>>> I also have the same feeling. Implementing in L2 may be
> >>>      >>     difficult to
> >>>      >>     >>> be used in e2e transport. Of course it can work well in
> >>>      limited
> >>>      >>     >>> domain, like DC or HPC clusters. However, I also look
> >>>      for some
> >>>      >>     >>> solutions that could be able to go through internet. We
> >>>      have
> >>>      >>     >>> submitted a draft to describe the transport challenges.
> See
> >>>      >>     >>>
> >>>      >> https://datatracker.ietf.org/doc/html/draft-huang-tsvwg-
> transport-
> >>>      >>     >>> challenges.
> >>>      >>     >>>>
> >>>      >>     >>>> I share the same opinion that the congestion signal is
> >>>      useful
> >>>      >> and
> >>>      >>     >>> current 1-bit ECN solution is not fully sufficient. But
> >>>      I also
> >>>      >>     feel
> >>>      >>     >>> like the more straight way is to extend L3, or l4, like
> >>>      update
> >>>      >>     IOAM,
> >>>      >>     >>> to carry the information. For L2 solution, it should be
> >>>      developed
> >>>      >>     >>> together with IEEE 802.1.
> >>>      >>     >>>>
> >>>      >>     >>>> BR，
> >>>      >>     >>>> Rachel
> >>>      >>     >>>>
> >>>      >>     >>>> 发件人: iccrg <[e-mail]> 代表 Tom Herbert
> >>>      >>     >>>> 发送时间: 2023年9月10日 0:10
> >>>      >>     >>>> 收件人: Abhiram Ravi
> >>>      <abhiramr=[e-mail]>
> >>>      >>     >>>> 抄送: IETF IPPM WG <[e-mail]>; tsvwg
> >>>      <[e-mail]>;
> >>>      >>     >>> [e-mail]; [e-mail]; Nandita Dukkipati
> >>>      >>     >>> <[e-mail]>; Naoshad Mehta
> >>>      <[e-mail]>; Jai
> >>>      >>     Kumar
> >>>      >>     >>> <[e-mail]>
> >>>      >>     >>>> 主题: Re: [iccrg] [tsvwg] New Internet Draft: Congestion
> >>>      >>     Signaling
> >>>      >>     >>> (CSIG)
> >>>      >>     >>>>
> >>>      >>     >>>> Hi, thanks for draft!
> >>>      >>     >>>>
> >>>      >>     >>>> The first thing that stands out to me is the carrier
> >>>      of the new
> >>>      >>     >>>> packet
> >>>      >>     >>> headers. In the forward path it would be in L2 and in
> >>>      >>     reflection it
> >>>      >>     >>> would be L4. As the draft describes, this would entail
> >>>      having to
> >>>      >>     >>> support the protocol in multiple L2 and multiple L4
> >>>      protocols--
> >>>      >>     >>> that's going to be a pretty big lift! Also, L2 is not
> >>>      really an
> >>>      >>     >>> end-to-end protocol (would legacy switches in the path
> also
> >>>      >>     forward the header)l?).
> >>>      >>     >>>>
> >>>      >>     >>>> The signaling being described in the draft is network
> >>>      layer
> >>>      >>     >>> information, and hence IMO should be conveyed in
> >>>      network layer
> >>>      >>     headers.
> >>>      >>     >>> That's is L3 which conveniently is the average of L2+L4
> :-)
> >>>      >>     >>>>
> >>>      >>     >>>> IMO, the proper carrier of the signal data is Hop-by-
> Hop
> >>>      >>     Options.
> >>>      >>     >>>> This
> >>>      >>     >>> is end-to-end and allows modification of data
> >>>      in-flight. The
> >>>      >>     typical
> >>>      >>     >>> concern with Hop-by-Hop Options is high drop rates on
> the
> >>>      >>     Internet,
> >>>      >>     >>> however in this case the protocol is explicitly
> >>>      confined to a
> >>>      >>     limited
> >>>      >>     >>> domain so I don't see that as a blocking issue for this
> >>>      use case.
> >>>      >>     >>>>
> >>>      >>     >>>> The information being carried seems very similar to
> >>>      that of IOAM
> >>>      >>     >>>> (IOAM
> >>>      >>     >>> uses Hop-by-Hop Options and supports reflection). I
> >>>      suppose the
> >>>      >>     >>> differences are that this protocol is meant to be
> >>>      consumed by the
> >>>      >>     >>> transport Layer and the data is a condensed summary of
> path
> >>>      >>     >>> characteristics. IOAM seems pretty extensible, so maybe
> it
> >>>      >>     could be
> >>>      >>     >>> adapted to carry the signals of this draft?
> >>>      >>     >>>>
> >>>      >>     >>>> A related proposal might be FAST draft-herbert-fast.
> Where
> >>>      >>     the CSIG
> >>>      >>     >>>> is
> >>>      >>     >>> network to host signaling, FAST is host to network
> >>>      signaling
> >>>      >>     for the
> >>>      >>     >>> purposes of requesting network services. These might be
> >>>      >>     complementary
> >>>      >>     >>> and options for both may be in the same packet. FAST
> >>>      also uses
> >>>      >>     >>> reflection, so we might be able to leverage some common
> >>>      >>     >>> implementation at a destination.
> >>>      >>     >>>>
> >>>      >>     >>>> Tom
> >>>      >>     >>>>
> >>>      >>     >>>> On Fri, Sep 8, 2023, 7:43 PM Abhiram Ravi
> >>>      >>     >>> <abhiramr=[e-mail]> wrote:
> >>>      >>     >>>> Hi IPPM folks,
> >>>      >>     >>>>
> >>>      >>     >>>> I am pleased to announce the publication of a new
> >>>      internet
> >>>      >> draft,
> >>>      >>     >>> Congestion Signaling (CSIG):
> >>>      >> https://datatracker.ietf.org/doc/draft-
> >>>      >>     >>> ravi-ippm-csig/
> >>>      >>     >>>>
> >>>      >>     >>>> CSIG is a new end-to-end packet header mechanism for
> >>>      in-band
> >>>      >>     >>>> signaling
> >>>      >>     >>> that is simple, efficient, deployable, and grounded in
> >>>      >>     concrete use
> >>>      >>     >>> cases of congestion control, traffic management, and
> >>>      network
> >>>      >>     >>> debuggability. We believe that CSIG is an important new
> >>>      >>     protocol that
> >>>      >>     >>> builds on top of existing in-band network telemetry
> >>>      protocols.
> >>>      >>     >>>>
> >>>      >>     >>>> We encourage you to read the CSIG draft and provide
> your
> >>>      >>     feedback
> >>>      >>     >>>> and
> >>>      >>     >>> comments. We have also cc'd the TSVWG, CCWG, and ICCRG
> >>>      mailing
> >>>      >>     lists,
> >>>      >>     >>> as we believe that this work may be of interest to their
> >>>      >>     members as well.
> >>>      >>     >>>>
> >>>      >>     >>>> Thank you for your time and consideration.
> >>>      >>     >>>>
> >>>      >>     >>>> Sincerely,
> >>>      >>     >>>> Abhiram Ravi
> >>>      >>     >>>> On behalf of the CSIG authors
> >>>      >>     >>
> >>>      >>     >
> >>>      >>     > _______________________________________________
> >>>      >>     > iccrg mailing list
> >>>      >>     > [e-mail]
> >>>      >>     > https://www.irtf.org/mailman/listinfo/iccrg
> >>>      >>
> >>>      >>     --     CCWG mailing list
> >>>      >> [e-mail]
> >>>      >> https://www.ietf.org/mailman/listinfo/ccwg
> >>>      >>
> >>>      >
> >>>      >
> >>>      > _______________________________________________
> >>>      > iccrg mailing list
> >>>      > [e-mail]
> >>>      > https://www.irtf.org/mailman/listinfo/iccrg
> >>>
> >>>      _______________________________________________
> >>>      iccrg mailing list
> >>>      [e-mail]
> >>>      https://www.irtf.org/mailman/listinfo/iccrg
> >>>
> >>>
> >>>
> >>> --
> >>> K. K. Ramakrishnan
> >>>
> >>> _______________________________________________
> >>> iccrg mailing list
> >>> [e-mail]
> >>> https://www.irtf.org/mailman/listinfo/iccrg
> >> --
> >> ________________________________________________________________
> >> Bob Briscoehttp://bobbriscoe.net/
> >> _______________________________________________
> >> iccrg mailing list
> >> [e-mail]
> >> https://www.irtf.org/mailman/listinfo/iccrg
> >
> 
> --
> ________________________________________________________________
> Bob Briscoe                               http://bobbriscoe.net/
> 
> 
>